C-GCN: A Flexible CSI Phase Feature Extraction Network for Error Suppression in Indoor Positioning.

Area of Science:

• Wireless Communications
• Indoor Positioning Systems
• Machine Learning for Signal Processing

Background:

• Channel State Information (CSI) is crucial for indoor positioning but is often distorted by environmental and hardware factors.
• Traditional methods struggle to extract reliable location features from multi-scene CSI phase data.
• Graph neural networks offer potential for indoor positioning but require specialized exploration.

Purpose of the Study:

• To propose a novel Cooperation-Graph Convolution Network (C-GCN) for enhanced indoor positioning.
• To extract new multi-dimensional correlation features from CSI phase data.
• To leverage graph neural networks for improved feature extraction in challenging indoor environments.

Main Methods:

• Developed a Cooperation-Graph Convolution Network (C-GCN) integrating convolution and graph convolution layers.
• Represented each antenna-subcarrier pair as a node in a graph, connecting them via correlation.
• Utilized graph convolution to aggregate node vectors and standard convolution to extract Euclidean space fluctuations from data packets.
• Applied end-to-end supervised training for effective feature extraction.

Main Results:

• The C-GCN demonstrated superior performance in restraining positioning errors.
• Achieved an average positioning error of 1.29 m in office environments and 1.71 m in garages.
• When combined with amplitude features, the average positioning error decreased to 0.99 m in offices and 1.14 m in garages.

Conclusions:

• The proposed C-GCN effectively extracts phase features from distorted CSI for accurate indoor positioning.
• The model shows significant improvements over existing methods in typical indoor environments.
• C-GCN offers a promising approach for robust and precise indoor localization using CSI.